if your getting very blurry outputs
IF YOUR ENCOUNTERING BLURRY/BAD IMAGES!
raise shift to 12-13! solves most issues with bad outputs!
Lighting lora compatible!
shift of 12-13
cfg of 1
8step lightning lora - 8 steps
V examples below V
shift - 12
shift -13
In your first samples, can you share which quant you used, and what your prompt/workflow is like? It sounds odd that cfg and steps wouldn't really have an impact so something else must be going on.
In your first samples, can you share which quant you used, and what your prompt/workflow is like? It sounds odd that cfg and steps wouldn't really have an impact so something else must be going on.
drag one into comfy and itll give you the entire workflow! <3 i believe the quant was the q4_0.
it was definately the shift being too low. not sure why because thats extremely unusual in my experience but ayy if it works, it works! i got it running proper at 12-13 shift
cfg of 4 made no sense either as gguf models usually have a strict cfg of 1 from past experience so i lowered that as well. lightning lora compatible as well so it drops step count down without issue for testing if you decide to run it.
In your first samples, can you share which quant you used, and what your prompt/workflow is like? It sounds odd that cfg and steps wouldn't really have an impact so something else must be going on.
drag one into comfy and itll give you the entire workflow! <3 i believe the quant was the q4_0.
it was definately the shift being too low. not sure why because thats extremely unusual in my experience but ayy if it works, it works! i got it running proper at 12-13 shift
Awesome, feel free to edit your post if you'd like to let people know if they're encountering any similar issue. We'll also put it in our guide so thanks for the feedback - very helpful :)
In your first samples, can you share which quant you used, and what your prompt/workflow is like? It sounds odd that cfg and steps wouldn't really have an impact so something else must be going on.
drag one into comfy and itll give you the entire workflow! <3 i believe the quant was the q4_0.
it was definately the shift being too low. not sure why because thats extremely unusual in my experience but ayy if it works, it works! i got it running proper at 12-13 shift
Awesome, feel free to edit your post if you'd like to let people know if they're encountering any similar issue. We'll also put it in our guide so thanks for the feedback - very helpful :)
happy to help! thank you so much for the model guys! calculator pc users like myself (rtx 3070) really appreciate all the work you guys put in. <3
Hi just to add I was able to get a decent result the first try using our workflow and this prompt
A beautiful Chinese woman wearing a T-shirt with the “Rebel AI” logo is holding a black marker and smiling at the camera. She is holding up a white sign that clearly reads “Qwen Image 2512”. The background is clean and minimal, with soft natural lighting and a professional, friendly vibe.
turbo lora 4steps
so im not sure what happened... this also happened when i first tested Flux 2, but the shift literally corrected itself. i started getting weird outputs so i lowered it back down to 4.0 and all the sudden i got clean outputs again lmao so idk maybe its a comfy issue that takes time to correct on its own. im confused but the 4.0 shift works as intended.
UPDATE!!
dype node works fairly well so far in testing! actually creates a very high fidelity image. some slight morphing and hallucination going on but that could be value-related. just wanted to let you guys know!
2160x1024
2160x1440 seems to be the correct landscape value
2160x1440
with just 10 steps. bmw m3 car.
